From c11062dd765d77f8e78ff5403541fe43085ad763 Mon Sep 17 00:00:00 2001
From: Volpeon <git@volpeon.ink>
Date: Fri, 13 Jan 2023 14:12:22 +0100
Subject: Simplified step calculations

---
 training/common.py | 42 ++++++++++++++++++++++--------------------
 1 file changed, 22 insertions(+), 20 deletions(-)

(limited to 'training')

diff --git a/training/common.py b/training/common.py
index 842ac07..180396e 100644
--- a/training/common.py
+++ b/training/common.py
@@ -36,21 +36,24 @@ def get_scheduler(
     warmup_exp: int,
     annealing_exp: int,
     cycles: int,
+    train_epochs: int,
     warmup_epochs: int,
     optimizer: torch.optim.Optimizer,
-    num_train_epochs: int,
-    num_update_steps_per_epoch: int,
+    num_training_steps_per_epoch: int,
     gradient_accumulation_steps: int,
 ):
-    num_train_steps = num_train_epochs * num_update_steps_per_epoch
-    warmup_steps = warmup_epochs * num_update_steps_per_epoch * gradient_accumulation_steps
+    num_training_steps_per_epoch = math.ceil(
+        num_training_steps_per_epoch / gradient_accumulation_steps
+    ) * gradient_accumulation_steps
+    num_training_steps = train_epochs * num_training_steps_per_epoch
+    num_warmup_steps = warmup_epochs * num_training_steps_per_epoch
 
     if id == "one_cycle":
         min_lr = 0.04 if min_lr is None else min_lr / lr
 
         lr_scheduler = get_one_cycle_schedule(
             optimizer=optimizer,
-            num_training_steps=num_train_steps * gradient_accumulation_steps,
+            num_training_steps=num_training_steps,
             warmup=warmup_func,
             annealing=annealing_func,
             warmup_exp=warmup_exp,
@@ -58,21 +61,21 @@ def get_scheduler(
             min_lr=min_lr,
         )
     elif id == "cosine_with_restarts":
-        cycles = cycles if cycles is not None else math.ceil(
-            math.sqrt(((num_train_steps - warmup_steps) / num_update_steps_per_epoch)))
+        if cycles is None:
+            cycles = math.ceil(math.sqrt(((num_training_steps - num_warmup_steps) / num_training_steps_per_epoch)))
 
         lr_scheduler = get_cosine_with_hard_restarts_schedule_with_warmup(
             optimizer=optimizer,
-            num_warmup_steps=warmup_steps,
-            num_training_steps=num_train_steps * gradient_accumulation_steps,
+            num_warmup_steps=num_warmup_steps,
+            num_training_steps=num_training_steps,
             num_cycles=cycles,
         )
     else:
         lr_scheduler = get_scheduler_(
             id,
             optimizer=optimizer,
-            num_warmup_steps=warmup_steps,
-            num_training_steps=num_train_steps * gradient_accumulation_steps,
+            num_warmup_steps=num_warmup_steps,
+            num_training_steps=num_training_steps,
         )
 
     return lr_scheduler
@@ -135,7 +138,7 @@ def loss_step(
     noise_scheduler: DDPMScheduler,
     unet: UNet2DConditionModel,
     text_encoder: CLIPTextModel,
-    num_class_images: int,
+    with_prior: bool,
     prior_loss_weight: float,
     seed: int,
     step: int,
@@ -184,7 +187,7 @@ def loss_step(
     else:
         raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
-    if num_class_images != 0:
+    if with_prior:
         # Chunk the noise and model_pred into two parts and compute the loss on each part separately.
         model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
         target, target_prior = torch.chunk(target, 2, dim=0)
@@ -226,11 +229,10 @@ def train_loop(
     on_after_optimize: Callable[[float], None] = noop,
     on_eval: Callable[[], _GeneratorContextManager] = nullcontext
 ):
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / gradient_accumulation_steps)
-    num_train_steps = num_epochs * num_update_steps_per_epoch
-
+    num_training_steps_per_epoch = math.ceil(len(train_dataloader) / gradient_accumulation_steps)
     num_val_steps_per_epoch = len(val_dataloader)
-    num_epochs = math.ceil(num_train_steps / num_update_steps_per_epoch)
+
+    num_training_steps = num_training_steps_per_epoch * num_epochs
     num_val_steps = num_val_steps_per_epoch * num_epochs
 
     global_step = 0
@@ -244,14 +246,14 @@ def train_loop(
     max_acc_val = 0.0
 
     local_progress_bar = tqdm(
-        range(num_update_steps_per_epoch + num_val_steps_per_epoch),
+        range(num_training_steps_per_epoch + num_val_steps_per_epoch),
         disable=not accelerator.is_local_main_process,
         dynamic_ncols=True
     )
     local_progress_bar.set_description(f"Epoch 1 / {num_epochs}")
 
     global_progress_bar = tqdm(
-        range(num_train_steps + num_val_steps),
+        range(num_training_steps + num_val_steps),
         disable=not accelerator.is_local_main_process,
         dynamic_ncols=True
     )
@@ -309,7 +311,7 @@ def train_loop(
 
                     local_progress_bar.set_postfix(**logs)
 
-                    if global_step >= num_train_steps:
+                    if global_step >= num_training_steps:
                         break
 
             accelerator.wait_for_everyone()
-- 
cgit v1.2.3-70-g09d2